Deriving ocean color products using neural networks

• Used novel neural network (NN) approach to retrieve ocean water components from Rrs
• Separated the non-phytoplankton absorption into CDOM and NAP components from Rrs
• Estimated chlorophyll concentration using derived IOPs and Rrs as inputs
• Validated all NNs using the synthetic datasets and field NOMAD dataset

In this paper we develop a neural network (NN) algorithm for retrieving inherent optical properties (IOP) from above water remote sensing reflectances (Rrs) at available MODIS (or similar satellite) wavelengths. In previous work we used Hydrolight5 simulations of Rrs with widely varying globally representative constituent physical parameters as a training basis to develop a neural network algorithm, which, using the Rrs at the MODIS visible wavelengths (412, 443, 488, 531, 547 and 667 nm) as input, retrieves the in-water particulate backscattering (bbp), phytoplankton (aph) and non-phytoplankton (adg) absorption coefficients at 443 nm. In this work, using the same dataset, we develop a NN which takes these same Rrs as input, and produces an output which is used to separate the non-phytoplankton absorption coefficient (adg) at 443 nm into dissolved (ag) and particulate (adm) components. We then apply this synthetically trained algorithm to the NASA bio-Optical Marine Algorithm Data set (NOMAD) Rrs to retrieve IOP at 443 nm, with the measured NOMAD and retrieved IOP values showing good agreement. These retrieved IOP along with their related Rrs values are then used to train an additional NN that produces chlorophyll concentration [Chl] as output. It is shown that these [Chl] values are retrieved more accurately when compared with ones retrieved with a similar approach which does not use IOP as input, as well as with those derived using the MODIS OC3 algorithm.